Traffic flow prediction system based on trip chain dynamic evolution

Abstract

The application relates to the technical field of intelligent transportation, and particularly discloses a traffic flow prediction system based on travel chain dynamic evolution. The system comprises travel chain real-time sensing and reconstruction modules, travel decision behavior dynamic deduction modules, group travel demand evolution calculation modules and macroscopic traffic flow dynamic prediction modules which are connected in sequence. Individual travel chains are sensed and reconstructed in real time, individual behaviors are dynamically deduced, group demands are aggregated based on agent simulation, and a macroscopic model is driven to perform prediction. The system builds a complete technical closed loop from real-time sensing of micro individual behaviors to dynamic prediction of macro road network flow. Through cooperation of the travel decision behavior dynamic deduction modules and the group travel demand evolution calculation modules, the system realizes fine simulation of group travel behavior emergent effect under a sudden situation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent transportation, and particularly relates to a traffic flow prediction system based on dynamic evolution of travel chains. BACKGROUND

[0002] Traffic flow prediction is a core and fundamental technology of intelligent transportation systems and urban traffic management. It analyzes historical and real-time traffic data to predict the traffic state of a specific period and road segment in the future, providing decision support for traffic induction, signal control, path planning, and other applications.

[0003] Modeling and prediction based on individual travel behavior is an important research direction to improve the accuracy of macroscopic traffic flow prediction. It aims to understand the generation and evolution mechanism of traffic demand by analyzing the complete activity sequence of travelers.

[0004] Existing technologies usually construct static or quasi-dynamic travel chain models based on historical data to describe the complete travel activity sequence of residents from the starting point to the destination within a day. However, such methods have significant limitations when facing sudden changes in traffic network state caused by traffic accidents, adverse weather, large-scale activities, etc.

[0005] The model parameter update of existing systems has a delay, making it difficult to capture real-time adjustments in traveler behavior caused by traffic accidents, such as route changes, trip cancellations, or activity sequence rearrangements. These minute-level dynamic evolutions cannot be reflected in the prediction model in a timely manner. Since the model fails to incorporate real-time evolving travel chain information, its prediction of traffic flow in subsequent periods will have systematic biases, leading to distorted prediction results and making it impossible to provide accurate and reliable basis for dynamic traffic management, thereby reducing the effectiveness of traffic emergency response and induction strategies. SUMMARY

[0006] The present application aims to provide a traffic flow prediction system based on dynamic evolution of travel chains to solve the problem that static or quasi-dynamic travel chain models in existing technologies cannot capture and respond to dynamic evolution of individual travel behavior caused by traffic accidents in real time, resulting in distorted macroscopic traffic flow prediction.

[0007] The technical solution of the present application is a traffic flow prediction system based on dynamic evolution of travel chains, which includes:

[0008] A travel chain real-time perception and reconstruction module for continuously accessing multi-source heterogeneous real-time traffic data streams and dynamically identifying and updating the state of active travel individuals in the region at a minute level based on the data.

[0009] The travel decision behavior dynamic deduction module is configured to receive the dynamic travel chain state output by the travel chain real-time perception and reconstruction module, and perform parallel deduction on the possible behavior of each travel individual at the next decision point based on a preset micro-behavior rule library and a real-time network state.

[0010] The group travel demand evolution calculation module is configured to receive all individual behavior deduction results output by the travel decision behavior dynamic deduction module, perform aggregation and interaction calculation based on an agent-based simulation framework, and generate a dynamic travel demand distribution of the whole road network in a future period.

[0011] The macro-traffic flow dynamic prediction module is configured to receive the dynamic travel demand distribution output by the group travel demand evolution calculation module, and use the dynamic travel demand distribution as an input boundary condition to drive a macro-traffic flow simulation model to run, and finally output a refined traffic flow and state prediction result of each road section and intersection in a specific future period.

[0012] Further, the travel chain real-time perception and reconstruction module comprises a data fusion sub-module, a chain identification sub-module and a state updating sub-module.

[0013] The data fusion sub-module is configured to access vehicle identity recognition data from a traffic checkpoint camera, mobile phone signaling positioning data from a mobile communication base station, trajectory point data from a floating car and accident and regulation information data from a traffic incident detection system in parallel; the sub-module is internally provided with a unified space-time reference alignment algorithm and a multi-source data confidence weighted fusion algorithm to generate a high-reliability individual space-time trajectory sequence.

[0014] The chain identification sub-module is configured to receive the fused individual space-time trajectory sequence, and apply a travel purpose and activity type identification algorithm based on a hidden Markov model to segment and label the continuous trajectory point sequence into two states of “travel” and “stay”, and then construct a complete “travel-stay-travel…” sequence as an initial travel chain; each travel chain comprises a chain identifier, an individual identifier, an activity sequence, a predicted end time of each activity and a current activity state.

[0015] The state updating sub-module is configured to monitor and update the current state of each active travel chain at a period of 1 minute; when it is detected that the real-time trajectory of an individual deviates from its preset path by more than a threshold distance, or the real-time stay time of the individual deviates from the predicted end time by more than a threshold proportion, or external event information affecting the current activity of the individual is received, the sub-module immediately triggers reconstruction of the travel chain, updates the subsequent activity sequence and time estimation of the travel chain, and marks the updated travel chain state as “evolved” and then outputs.

[0016] Further, the travel decision behavior dynamic deduction module comprises a rule engine sub-module, a network state interface sub-module and a parallel deduction sub-module.

[0017] a rule engine sub-module, which is built-in with a structured microscopic travel behavior rule base that defines the behavior options and their selection probabilities that an individual might take under different triggering conditions; the rule conditions include the current activity type, time urgency, real-time road condition information, and individual historical preference pattern; the behavior options include continuing the original plan, changing the path, canceling the next activity, adjusting the activity sequence, and switching the travel mode.

[0018] a network state interface sub-module, which is used to obtain the traffic state index, road section travel time, signal timing scheme, and event influence range map of the entire road network in real time, and convert them into standardized input parameters that can be called by the rule engine.

[0019] a parallel deduction sub-module, which is used to create independent deduction instances for each travel chain marked as "evolved" or approaching a decision point; each instance loads the corresponding individual attributes and current chain state, combines the real-time parameters provided by the network state interface sub-module, drives the rule engine to make logical judgments, calculates the most likely behavior and its probability that the individual will take at the next decision point, and outputs structured behavior deduction instructions.

[0020] Further, the group travel demand evolution calculation module adopts an agent-based simulation framework, which includes an agent encapsulation unit, an environment modeling unit, and an interactive calculation unit.

[0021] The agent encapsulation unit is used to encapsulate each travel chain and its corresponding dynamic deduction instructions into a travel agent with autonomous decision-making capability; each agent has attributes such as location, destination, travel mode, path selection strategy, and behavior rules.

[0022] The environment modeling unit is used to build a digital twin road network environment, which includes road network topology, road section capacity, signal control logic, and current traffic load conditions fed back by the macroscopic traffic flow dynamic prediction module.

[0023] The interactive calculation unit is used to execute the decision-making and movement of all travel agents in parallel within a set simulation time step; agents act in the environment according to their deduction instructions, and their path selection is influenced by real-time road conditions and other agent decisions; through iterative calculation, the game and cooperation between individuals are simulated; this unit finally counts the origin-destination matrix and path flow generated by all travel agents on each road segment in the road network within each simulation step, and aggregates to generate the dynamically changing travel demand distribution in the future period.

[0024] Further, the macroscopic traffic flow dynamic prediction module includes a model coupling interface sub-module, a macroscopic simulation sub-module, and a result output sub-module.

[0025] A model coupling interface submodule is configured to receive dynamic travel demand distribution data output by the group travel demand evolution calculation module, and convert the dynamic travel demand distribution data into an input file format required by a macro traffic flow simulation model according to a time slice, so as to ensure that the demand data is accurately aligned with a simulation time window.

[0026] A macro simulation submodule is configured to use a dynamic traffic distribution model or a cell transmission model as a core simulation engine; the engine takes an actual traffic flow state at a current time as an initial condition, and takes dynamic travel demand distribution as an input, simulates a convergence, dissipation and queuing process of vehicles in a road network, and calculates a vehicle flow, a density, a speed and a delay of an intersection of each road segment in each simulation period within 5 minutes to 2 hours in the future.

[0027] A result output submodule is configured to perform post-processing on original results output by the macro simulation submodule, including data smoothing, abnormal value correction and visualization packaging, and finally generate a traffic flow prediction report, a state level diagram and congestion warning information which are subdivided according to road segments and time intervals.

[0028] Further, in the travel chain real-time perception and reconstruction module, weight coefficients of the multi-source data confidence weighted fusion algorithm are dynamically adjusted according to real-time coverage, historical accuracy and inherent bias in a specific space-time scene of the data source; the adjustment period is 15 minutes, and the error statistical results of each data source compared with high-precision reference data in the last period are used for updating.

[0029] Further, the micro travel behavior rule library in the travel decision behavior dynamic deduction module supports online updating based on a reinforcement learning mechanism; the system continuously collects differences between actual behavior results and deduced predicted behaviors of individuals, and uses the differences as feedback signals to fine-tune selection probability parameters of corresponding rules at a learning rate of 0.01, so that the rule library can adaptively track long-term slow evolution of group behavior patterns.

[0030] Further, a closed-loop feedback link is formed between the group travel demand evolution calculation module and the macro traffic flow dynamic prediction module; the macro traffic flow dynamic prediction module feeds back the predicted future road condition state to an environment modeling unit of the group travel demand evolution calculation module in advance by a simulation period, so as to update a network state perceived by an intelligent agent when making a decision, thereby more realistically simulating a dynamic interaction process between travel demand and traffic supply.

[0031] Further, the entire system runs under a unified time management framework, which synchronizes real-time flow of a physical world and deduction flow of a simulation world; the system sets a rolling prediction window with a length of 10 minutes, performs full-process calculation from perception, deduction, evolution to prediction every 1 minute, and outputs updated prediction results, so as to realize continuous, rolling and refined prediction of a future traffic state.

[0032] The beneficial effects of the present application compared with the prior art are:

[0033] 1. The present application constructs a complete technical closed loop from real-time perception of microscopic individual behavior to macroscopic road network traffic dynamic prediction. Through the travel chain real-time perception and reconstruction module, the core contradiction that the traditional static model cannot respond to real-time behavior changes is fundamentally solved. This bottom-up modeling method makes the source data of traffic demand prediction always synchronized with the actual situation, significantly improving the timeliness of the prediction model and the authenticity of the basic data.

[0034] 2. The present application realizes fine simulation of the emergent effect of group travel behavior under sudden situations through the cooperation of the travel decision behavior dynamic deduction module and the group travel demand evolution calculation module. The system not only considers the direct reaction of individuals to events, but also simulates the mutual influence between individual decisions and road condition feedback based on agent, which can predict the secondary congestion or traffic transfer phenomenon caused by group path switching. This ability to describe the cascading effect of complex systems makes the prediction results more close to the dynamic evolution law of real traffic network.

[0035] 3. The present application adopts a hierarchical decoupling, modularized system architecture, and introduces a closed-loop feedback and adaptive learning mechanism. The function boundaries of each module are clear, coupled through standardized data interfaces, which not only ensures the efficient operation and scalability of the whole system, but also through the feedback between demand evolution and flow prediction, online learning of behavior rule library, gives the system continuous self-optimization ability. The system can not only cope with sudden traffic abnormalities, but also gradually adapt to the long-term changes of urban travel patterns, providing a precise, reliable, robust and growing core prediction engine for intelligent traffic management. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is the overall technical scheme architecture diagram of the traffic flow prediction system based on travel chain dynamic evolution proposed by the present application;

[0037] Figure 2 is the core principle framework diagram of the travel chain real-time perception and reconstruction module in the present application;

[0038] Figure 3 is the collaborative logic flow framework diagram of the travel decision behavior dynamic deduction module and the group travel demand evolution calculation module in the present application;

[0039] Figure 4 is the closed-loop feedback and data flow diagram between the group travel demand evolution calculation module and the macroscopic traffic flow dynamic prediction module in the present application;

[0040] Figure 5is the core principle of the present application on the rolling prediction mechanism and the prior art comparison diagram. DETAILED DESCRIPTION

[0041] The present application provides a traffic flow prediction system based on the dynamic evolution of travel chains. Please refer to the attached Figure 1 to the attached Figure 5 , the system constructs a complete technical closed loop from real-time perception of micro individual behavior to dynamic prediction of macro road network traffic. The core of the system consists of four modules in sequence: travel chain real-time perception and reconstruction module, travel decision behavior dynamic deduction module, group travel demand evolution calculation module and macro traffic flow dynamic prediction module.

[0042] The whole system runs under a unified time management framework, which strictly synchronizes the real-time data stream of the physical world with the deduction calculation stream of the simulation world. The system sets a rolling prediction window of 10 minutes, and triggers the whole process calculation from data perception, behavior deduction, demand evolution to macro prediction once every 1 minute, and finally outputs the updated fine traffic flow and state prediction results within 10 minutes to 2 hours in the future. This rolling execution mechanism ensures that the prediction results can continuously track the dynamic evolution of the traffic system, please refer to the attached Figure 5 , which shows the advantages of the rolling prediction mechanism of the present application compared with the traditional static or long cycle batch prediction method.

[0043] Firstly, the travel chain real-time perception and reconstruction module is responsible for continuously accessing multi-source heterogeneous real-time traffic data streams, and dynamically identifying and updating the complete travel chain of active travel individuals in the region with minute-level granularity. This module is the data source and situation awareness basis of the whole system. Please refer to the attached Figure 2 , the travel chain real-time perception and reconstruction module is further composed of three sub-modules: data fusion sub-module, chain identification sub-module and state update sub-module, which work together to realize the transformation from raw data to structured and updateable travel chain state.

[0044] The data fusion sub-module is the multi-source data entrance and preprocessing center of the system. This sub-module accesses real-time data streams from four main channels in parallel.

[0045] The first road data comes from the city traffic portal camera system, which captures vehicle images passing through the portal at a frequency of several frames per second, and extracts vehicle identity, passing time and geographic location through license plate recognition algorithm.

[0046] The second road data comes from mobile communication operators, and the system obtains anonymized mobile phone signaling data through a special data interface in real time. These data record the time stamp and base station geographic location code when the mobile phone and the base station interact with the signaling, which can be used to reverse the coarse-grained mobile trajectory of the individual.

[0047] The third data comes from floating cars equipped with global positioning system terminals, including taxis, buses and some freight vehicles, which report high-precision trajectory points containing latitude, longitude, instantaneous speed and direction angle at intervals of 10 to 30 seconds.

[0048] The fourth data comes from traffic incident detection systems and traffic control information release platforms, providing the occurrence time, location, impact range and expected duration of traffic accidents, road construction and large-scale event control in the form of structured messages.

[0049] In the face of these multi-source data with significant differences in time resolution, spatial accuracy, coverage range and reliability, the core task of the data fusion sub-module is to generate high-reliability individual spatio-temporal trajectory sequences. This task is achieved through two core algorithms. First is the unified spatio-temporal reference alignment algorithm. The timestamps of all input data are converted to millisecond-level timestamps based on coordinated universal time.

[0050] All geographic location information, whether latitude and longitude coordinates, base station numbers or toll gate numbers, are unified to planar coordinates in the same high-precision urban geographic information system coordinate system and associated to specific road segments and intersection numbers. Second is the multi-source data confidence weighted fusion algorithm. This algorithm dynamically assigns a confidence weight coefficient to each data source in each calculation period. The determination of the weight coefficient is based on three dimensions of evaluation: the real-time coverage rate of the data source, i.e. the proportion of individuals that the data source can provide effective data in the current spatio-temporal unit; the historical accuracy rate of the data source, i.e. the average error after comparing the data provided by the data source with high-precision reference data such as differential global positioning system trajectory in the past period of time; and the inherent bias in a specific spatio-temporal scenario, for example, mobile phone signaling data will be invalid in tunnels, while global positioning system data will have increased error in high-rise building dense areas.

[0051] The system takes 15 minutes as the evaluation period, dynamically updates the weight coefficient table for the next period based on the error statistical results of each data source compared with the high-precision reference data in the last period. During fusion, for multiple location estimates of the same individual at the same time, the data fusion sub-module performs weighted averaging according to the real-time weight of each data source and eliminates abnormal points that deviate significantly from other majority source data, finally outputting a time-continuous, spatially-smooth and confidence-evaluated individual spatio-temporal trajectory sequence.

[0052] The chain recognition sub-module receives the cleaned and fused individual spatio-temporal trajectory sequences from the data fusion sub-module. Its core function is to apply a travel purpose and activity type recognition algorithm based on Hidden Markov Model to transform the continuous trajectory point sequence into structured travel chains. The algorithm models the individual's movement process as a transition process between a series of hidden states, which represent different activity types, such as home, work, shopping, dining, etc., and travel states. The observation values are the spatial location, stay duration, arrival time, etc. of the trajectory points.

[0053] The algorithm obtains the state transition probability matrix and observation probability matrix through a large amount of historical data training. In real-time recognition, the chain recognition sub-module performs sliding window analysis on each individual's trajectory sequence, calculates the probability of each trajectory point belonging to a travel state or a stay state, and clusters the continuous points with a probability exceeding 0.95 into a state segment. A complete trip is defined as the movement process from the end point of a stay segment to the start point of the next stay segment. By decoding the most likely hidden state sequence through the Viterbi algorithm, the system constructs a complete trip stay sequence into an initial travel chain.

[0054] Each travel chain is represented as a structured data object in the system, and its key fields include: a globally unique chain identifier, an individual identifier for associating individuals and chains, an activity sequence arranged in chronological order, each activity sequence containing activity type code, activity location coordinates, expected start time and expected end time, and a field marking the current activity state, which can take values such as in progress, completed, canceled or pending evolution.

[0055] The state update sub-module is the key to ensuring the dynamic nature of travel chains. This sub-module performs state monitoring and updating on all active travel chains marked as in progress at a fixed period of 1 minute. Monitoring is triggered based on three parallel rules. The first is path deviation monitoring. The system calculates an expected path based on the start and end points of the current travel activity in the travel chain, combined with the real-time road network status, through a path planning algorithm. The state update sub-module continuously compares the nearest distance between the individual's real-time trajectory points and the expected path. When the average deviation distance of three consecutive trajectory points exceeds the preset threshold of 200 meters, it is determined that there is a significant deviation in the path. The second is stay duration deviation monitoring.

[0056] For the trip chains in stay state, the system compares the actual stay duration of the individual with the difference between the end time and start time of the activity. When the actual stay duration exceeds 150% of the expected duration or is less than 50%, it is determined that the activity duration deviates significantly. The third rule is external event impact monitoring. The state update submodule monitors the event information stream from the data fusion submodule in real time. When a new event is received, the system immediately calculates the geographical location and impact range polygon of the event, and determines whether the individual in the current trip chain is located within the polygon or the planned path, or whether the event type directly affects the current activity type.

[0057] Once any of the above rules is triggered, the state update submodule immediately starts the reconstruction process for the trip chain. Reconstruction is not a complete identification from scratch, but a partial correction based on the existing chain structure. For example, for path deviation, the system will try to identify the individual's new possible destination and update the subsequent activity sequence; for stay duration deviation, the system will adjust the expected end time of the current activity and may rearrange the timing of subsequent activities according to time constraints; for event impact, the system will insert a pending decision item to change the path or cancel the activity for the trip chain according to the event type under the guidance of the rule library.

[0058] After reconstruction is completed, the state update submodule marks the updated trip chain state as evolved and sends it as output to the downstream trip decision behavior dynamic inference module. Trip chains that do not evolve are only updated for their current activity progress and are not marked as evolved, but are also periodically output for the downstream module to perceive the overall trip situation.

[0059] Secondly, the trip decision behavior dynamic inference module receives the trip chain state from the upstream that is marked as evolved or is about to reach the preset decision point. Its core task is to parallelly infer the possible behavior of each trip individual at the next decision point based on the preset micro-behavior rules and real-time network state, and generate structured behavior instructions. Please refer to the attached Figure 3 The module works collaboratively with three submodules: rule engine submodule, network state interface submodule, and parallel inference submodule.

[0060] The rule engine submodule embeds a structured micro-trip behavior rule library. The rule library is organized using a production rule system, and each rule consists of a condition part and a behavior part. The condition part is a combination of logical expressions, which can include variables such as current activity type, e.g. work, leisure, commute; time urgency, calculated based on the difference between current time and the next activity expected start time; real-time traffic information, such as the travel time increment of the current path; and individual historical preference patterns, such as the individual's preference for highways during similar evening peak hours.

[0061] The behavior part defines one or more behavior options that the individual might take under the condition combination, and assigns an initial selection probability to each option. The main behavior options include: continue the original plan, i.e., not change any subsequent schedule; change route, i.e., replace the travel route from the current location to the next destination; cancel the next activity, i.e., give up the next stay activity in the original plan and go back or go to another place directly; adjust activity order, i.e., exchange the execution order of the next two or more activities; switch travel mode, e.g., from driving to subway. The rule base supports online updating based on reinforcement learning mechanism. The system continuously collects the actual behavior results of the individual, i.e., the difference between the real action of the individual observed by the state update sub-module and the behavior instruction predicted by the inference module before.

[0062] The difference is quantified as a reward or punishment signal. The system uses the stochastic gradient descent method to fine-tune the probability parameters of the corresponding behavior options that trigger the rules with a learning rate of 0.01. For example, if the rule predicts that a certain type of individual has a 70% probability of changing the route when encountering a certain type of accident, but the actual observation data shows that 85% of the individuals have changed the route, the probability of changing the route in this rule will be slightly adjusted upwards in subsequent calculations. This mechanism enables the rule base to adaptively track the long-term slow evolution of group behavior patterns.

[0063] The network state interface sub-module is responsible for providing the real-time environmental parameters required for decision-making for the rule engine. This sub-module obtains dynamic information of the entire road network from the city traffic brain or real-time traffic flow simulation model at a frequency of 1 time per second through a dedicated data bus. These information includes: traffic state index represented by color level, reflecting the congestion degree of road segment; real-time travel time of each road segment calculated based on floating car data; current timing plan of intersection signal lights and planned scheme of next several cycles obtained from signal control system; and dynamic change map of event influence range with influence intensity decay model obtained from event handling system.

[0064] The network state interface sub-module standardizes these heterogeneous information and converts them into a series of key-value pairs or vector-form parameters, such as converting road segment travel time into delay ratio relative to free flow time, and converting event influence map into spatial risk probability value related to individual route. These standardized parameters are cached in the in-memory database for low-latency query and invocation by the rule engine sub-module.

[0065] The parallel inference sub-module is the executor of the inference task. When receiving a batch of travel chain states that need to be inferred, this sub-module uses thread pool or distributed computing framework to create independent inference instances for each travel chain. Each instance loads the corresponding travel chain data object, including individual attributes and current chain state.

[0066] Then, the instance determines the next decision point according to the stage of the trip chain, for example, when the current activity is about to end, or when an incident is detected ahead of the path. Next, the reasoning instance pulls real-time network parameters related to the decision point from the network state interface submodule, such as the current travel time of each segment on the individual's planned path, the travel time of alternative paths, the impact range of the incident, etc.

[0067] Finally, the reasoning instance drives the rule engine submodule, taking the trip chain state and real-time network parameters as inputs, to match the conditions in the rule library. The rule engine performs forward reasoning to calculate the comprehensive probability of each behavior option in all matching rules. The system usually selects the behavior option with the highest comprehensive probability as the most likely behavior of the individual. The reasoning result is encapsulated as a structured behavior reasoning instruction object, whose fields include: the associated chain identifier, the decision timestamp, the reasoned behavior type, the behavior parameters, and the confidence probability of the behavior. The reasoning instructions of all individuals are batched and output to the downstream group travel demand evolution calculation module.

[0068] Third, the group travel demand evolution calculation module receives all individual behavior reasoning instructions from the travel decision behavior dynamic reasoning module. Its core task is to simulate the interaction and game of a large number of individuals in the road network environment through an agent-based simulation framework, and aggregate to generate the dynamic travel demand distribution of the entire road network in the future period. Please refer to the attached Figure 3 With the attached Figure 4 This module uses an agent-based simulation framework, which consists of three parts: agent encapsulation unit, environment modeling unit, and interaction calculation unit.

[0069] The agent encapsulation unit is responsible for instantiating the abstract trip chain and behavior instruction into a travel agent with autonomous decision-making ability. This unit creates an agent object for each trip chain carrying a reasoning instruction. Each agent object inherits all the properties of its trip chain, including the current location, destination sequence, travel mode, and additionally encapsulates the behavior strategy determined by the reasoning instruction, such as the path selection strategy. The path selection strategy may be based on the change path behavior in the reasoning instruction, and can be specified as a real-time dynamic path planning algorithm that aims to minimize the generalized travel cost, considering real-time travel time, toll fees, individual preferences, etc. in the cost function. The agent encapsulation unit assigns a unique agent identifier to all agents and establishes a mapping relationship between the original trip chain identifier and the agent identifier for data tracing.

[0070] The environment modeling unit is responsible for building a high-fidelity digital twin road network environment as the stage for the activities of agents. The environment is built based on the road network data of the city geographic information system and is a topological graph containing nodes and directed edges. Each node represents an intersection or an important transportation hub and contains coordinate and type attributes. Each directed edge represents a road segment and contains static attributes such as length, number of lanes, design speed, capacity, slope, and curvature. More importantly, the environment modeling unit dynamically injects traffic state information.

[0071] This information comes from two aspects: on the one hand, the current real-time traffic from the network state interface submodule, as the initial background traffic for the simulation; on the other hand, and also the key to the implementation of the closed-loop feedback of the system, the future predicted traffic from the macroscopic traffic flow dynamic prediction module. Please refer to the attached Figure 4 The macroscopic traffic flow dynamic prediction module feeds back the predicted future road segment travel time, flow, and other states to the environment modeling unit in advance of the simulation period. The environment modeling unit updates the road segment travel cost in the digital twin road network for the corresponding future time based on the feedback time tag, so that the agents can perceive the future traffic changes that may be caused by their own and other agents' decisions when making path decisions, thereby simulating more realistic prediction and game behaviors.

[0072] The interaction computing unit is the core engine of the simulation execution. The unit sets a fixed simulation time step, for example, 6 seconds. Within the simulation step, the interaction computing unit parallelly executes the decision and movement logic of all active agents. The decision logic triggers according to the state of the agent: the agent in the travel state decides the direction of the next movement or whether to trigger path re-planning based on its path selection strategy and the current perceived road network environment cost; the agent in the stay state checks its stay timer to determine whether it has reached the departure time. The movement logic calculates the distance that the agent can advance within the step based on the path selected by the agent and the road segment attributes, and updates its position.

[0073] All agent movements are parallelly computed but sequentially updated. The system uses a lock mechanism or a parallel conflict detection algorithm to handle the resource competition problem of agents at bottlenecks such as intersections, simulating microscopic behaviors such as car following, lane changing, and queuing. Through thousands of iterations of time steps, the system simulates the complete activity process of all traveling agents in the road network for several hours starting from the current time. The interaction computing unit at the end of each simulation step, statistics the travel demand generated by all agents: records the origin-destination information of each agent, as well as its traffic contribution on each road segment.

[0074] By accumulating the contribution of all agents and aggregating according to the actual time, for example, aggregating once every 5 minutes, the dynamically changing travel demand distribution data in the future period is finally generated. This data is usually represented as a four-dimensional tensor: origin partition, destination partition, travel mode, time slice, and the value of each unit represents the travel volume from a certain origin to a certain destination using a certain mode in that time slice.

[0075] Fourthly, the macroscopic traffic flow dynamic prediction module receives the dynamic travel demand distribution data from the group travel demand evolution calculation module. Its core task is to convert this demand expressed in the form of travel origin-destination into a refined traffic flow and state prediction of specific road segments and intersections on the road network. Please refer to the attached Figure 4 , which includes a model coupling interface submodule, a macroscopic simulation submodule, and a result output submodule.

[0076] The model coupling interface submodule is responsible for data format conversion and time alignment. Since the macroscopic traffic flow simulation model usually requires input files in a specific format, such as representing travel demand as a specific matrix, the model coupling interface submodule will slice and convert the received dynamic travel demand distribution tensor into the required input file sequence according to the preset time slice length, for example, 5 minutes. The conversion process ensures that the timestamps of the demand data are accurately aligned with the start time and simulation period of the macroscopic simulation model. At the same time, this submodule is also responsible for obtaining the actual traffic flow, average speed, occupancy rate, and other state data of each road segment and intersection on the entire road network at the current time from the real-time traffic flow database, serving as the initial conditions for the macroscopic simulation model.

[0077] The macroscopic simulation submodule is the core engine for flow calculation. In this embodiment, the cellular transmission model is used as the core simulation engine. The cellular transmission model discretizes each lane in the urban road network into a series of connected cells, and each cell can only accommodate a limited number of vehicles within a single simulation step. The state transition of the model is described by two basic equations.

[0078] The first is the conservation equation, which represents the change in the number of vehicles in the cell as the difference between the inflow and outflow. The second is the transmission equation, which specifies the maximum number of vehicles flowing from one cell to the next, which is determined by the current number of vehicles in the sending cell, the remaining capacity of the receiving cell, and the bottleneck road capacity connecting the two cells. Its core relationship can be expressed as:

[0079] where, represents the number of vehicles transmitted from cell i to cell in time step t;

[0080] represents the number of vehicles in cell i at the beginning of time step t; is the maximum outflow rate of cell i; is the capacity of cell ;

[0081] is the time step, tstartis the number of vehicles in cell at the beginning of time step t; is the congestion propagation parameter.

[0082] The macroscopic simulation submodule takes the actual traffic flow state at the current time as the initial condition, and injects the dynamic travel demand distribution into the corresponding origin cell of the road network in the form of vehicle generation. The model iteratively runs according to the set simulation step size, for example, 6 seconds, simulating the convergence, dissipation and queuing process of vehicles in the road network according to the cell transmission rules. Through hundreds of step calculations, the model can deduce the vehicle flow, vehicle density, space average speed of each simulation period of each cell of each road segment, and the queuing length and delay time of each approach of the intersection within 5 minutes to 2 hours in the future.

[0083] The result output submodule is responsible for post-processing and productization of the massive raw data output by the macroscopic simulation submodule. Post-processing includes data smoothing, using a moving average filter to eliminate jagged noise caused by simulation step size fluctuations; outlier correction, identifying and correcting obviously unreasonable data points caused by instantaneous mismatch of model parameters based on historical statistical rules.

[0084] Productization is to encapsulate the processed data into different forms of output products. The main products include: traffic flow prediction reports subdivided by road segment number and time interval, which can contain predicted flow, predicted speed, and predicted congestion index; traffic state level map based on geographic information system, which renders the road conditions at different time periods with different colors; and congestion warning information, the system generates different levels of warning information according to the degree and duration of predicted queuing length or delay time exceeding the threshold, and identifies the congestion start time, expected duration and impact range.

Claims

1. A traffic flow prediction system based on dynamic evolution of travel chains, characterized in that, include: The real-time perception and reconstruction module of the travel chain is used to continuously access multi-source heterogeneous real-time traffic data streams, and based on this, to perform minute-level dynamic identification and status updates of the travel chains of active individuals in the region. The travel decision-making behavior dynamic simulation module is used to receive the dynamic travel chain status output by the travel chain real-time perception and reconstruction module, and to perform parallel simulation of the possible behaviors of each travel individual at the next decision point based on the preset micro-behavioral rule base and real-time network status. The group travel demand evolution calculation module is used to receive all individual behavior inference results output by the travel decision behavior dynamic inference module, and aggregate and interact through an agent-based simulation framework to generate the dynamic travel demand distribution of the entire road network in the future period. The macro-traffic flow dynamic prediction module is used to receive the dynamic travel demand distribution output by the group travel demand evolution calculation module, and use it as input boundary conditions to drive the macro-traffic flow simulation model to run, and finally output the refined traffic flow and state prediction results of each road segment and intersection in a specific future time period. The real-time perception and reconstruction module of the travel chain includes a data fusion submodule, a chain recognition submodule, and a status update submodule. The data fusion submodule is used to access vehicle identification data from traffic checkpoint cameras, mobile phone signaling and positioning data from mobile communication base stations, trajectory point data from floating cars, and accident and control information data from traffic incident detection systems in parallel. This data fusion submodule incorporates a unified spatiotemporal benchmark alignment algorithm and a multi-source data confidence weighted fusion algorithm to generate highly reliable individual spatiotemporal trajectory sequences. The chain recognition submodule is used to receive the individual spatiotemporal trajectory sequence generated by the data fusion submodule, and apply the travel purpose and activity type recognition algorithm based on the hidden Markov model to segment and label the continuous trajectory point sequence into two states: travel and stay, and then construct an initial travel chain from a complete travel-stay-travel sequence. Each travel chain includes a chain identifier, an individual identifier, an activity sequence, the estimated end time of each activity, and the current activity status; The status update submodule is used to monitor and update the current status of each active travel chain at a 1-minute interval. When an individual's real-time trajectory deviates from its preset path by more than a threshold distance, or its real-time stay time deviates from the expected end time by more than a threshold ratio, or when external event information affecting its current activity is received, the status update submodule immediately triggers the reconstruction of the travel chain, updates its subsequent activity sequence and time estimate, and marks the updated travel chain status as evolved before outputting it. The dynamic simulation module for travel decision-making behavior includes a rule engine submodule, a network status interface submodule, and a parallel simulation submodule. The rule engine submodule has a built-in structured micro-travel behavior rule library, which defines the possible behavioral options and their probabilities for individuals under different triggering conditions. The rules and conditions include the current activity type, time urgency, real-time traffic information, and individual historical preference patterns; The behavior options include continuing the original plan, changing the path, canceling the next activity, adjusting the activity sequence, and switching the travel mode; The network state interface submodule is configured to acquire traffic state indexes, road section travel times, signal timing schemes, and event influence range maps of the whole road network in real time, and convert the indexes, times, schemes, and maps into standardized input parameters that can be called by the rule engine submodule; The parallel deduction submodule is configured to create independent deduction instances for each travel chain marked as evolved or adjacent to a decision point; Each instance loads corresponding individual attributes and a current chain state, combines real-time parameters provided by the network state interface submodule, drives the rule engine submodule to perform logical judgment, and calculates a behavior most likely to be taken by the individual at a next decision point and a probability of the behavior, and outputs a structured behavior deduction instruction; The group travel demand evolution calculation module adopts an agent-based simulation framework, which includes an agent encapsulation unit, an environment modeling unit, and an interaction calculation unit. The agent encapsulation unit is configured to encapsulate each travel chain and a corresponding dynamic deduction instruction as a travel agent with autonomous decision-making capability. Each agent has location, destination, travel mode, path selection strategy, and behavior rule attributes. The environment modeling unit is configured to build a digital twin road network environment, which includes a road network topology, a road section traffic capacity, a signal control logic, and a current traffic load condition fed back by the macroscopic traffic flow dynamic prediction module. The interaction calculation unit is configured to perform decision-making and movement of all travel agents in parallel within a set simulation time step. The agents act in the environment according to their deduction instructions, and their path selection is influenced by real-time road conditions and decisions of other agents. Through iterative calculation, the game and cooperation between individuals are simulated. The interaction calculation unit finally counts travel origin-destination matrices and path flows of all travel agents on each road section of the road network within each simulation step, and aggregates the matrices and flows to generate a dynamically changing travel demand distribution in a future time period. The macroscopic traffic flow dynamic prediction module includes a model coupling interface submodule, a macroscopic simulation submodule, and a result output submodule. The model coupling interface submodule is configured to receive dynamic travel demand distribution data output by the group travel demand evolution calculation module, and convert the data into an input file format required by a macroscopic traffic flow simulation model according to time slices, so as to ensure accurate alignment of the demand data and a simulation time window. The macroscopic simulation submodule adopts a dynamic traffic assignment model or a cell transmission model as a core simulation engine. The core simulation engine takes an actual traffic flow state at a current time as an initial condition, and takes a dynamic travel demand distribution as an input, simulates the convergence, dissipation, and queuing process of vehicles in the road network, and calculates vehicle flows, densities, speeds, and intersection delays of each simulation period of each road section in the future 5 minutes to 2 hours. The result output submodule is configured to post-process original results output by the macroscopic simulation submodule, including data smoothing, abnormal value correction, and visualization encapsulation, and finally generates a traffic flow prediction report, a state level map, and congestion warning information that are subdivided according to road sections and time intervals.

2. The system for traffic flow prediction based on dynamic evolution of travel chains according to claim 1, wherein, The weight coefficient of the multi-source data confidence weighted fusion algorithm is dynamically adjusted according to real-time coverage, historical accuracy and inherent deviation in a specific space-time scene of the data source; The adjustment period is 15 minutes, and the error statistical results of each data source compared with high-precision reference data in the last period are updated.

3. The system for traffic flow prediction based on dynamic evolution of travel chains according to claim 2, wherein, The microscopic travel behavior rule library supports online updating based on a reinforcement learning mechanism; The system continuously collects the difference between the actual behavior results of individuals and the deduced predicted behaviors, and takes this difference as a feedback signal to fine-tune the selection probability parameters of the corresponding rules at a learning rate of 0.

01.

4. The system for traffic flow prediction based on dynamic evolution of travel chains according to claim 3, wherein, A closed-loop feedback link is formed between the group travel demand evolution calculation module and the macroscopic traffic flow dynamic prediction module. The macroscopic traffic flow dynamic prediction module feeds back the predicted future road condition state to the environment modeling unit of the group travel demand evolution calculation module in advance by a simulation period, for updating the network state perceived by the agent when making decisions.

5. The system for traffic flow prediction based on dynamic evolution of travel chains according to claim 4, wherein, The core transmission rule of the cell transmission model adopted by the macroscopic simulation submodule is as follows: within a single simulation step, the number of vehicles transmitted from one cell to the next cell is determined by the current number of vehicles in the sending cell, the maximum outflow rate of the sending cell, the remaining capacity of the receiving cell and the bottleneck road section passing capacity connecting the two cells, and the minimum value of the above values is taken as the actual number of transmitted vehicles.

6. The system for traffic flow prediction based on dynamic evolution of travel chains according to claim 5, wherein, The entire system runs under a unified time management framework that synchronizes the real-time flow of the physical world with the deduced flow of the simulation world. The system sets a rolling prediction window of 10 minutes, and performs the whole-process calculation from perception, deduction, evolution to prediction every 1 minute, and outputs the updated prediction results.

Images